Neural network frequency response model and method based on physical information
By using a frequency response model based on a physical information neural network, the problem of traditional models being unable to quickly and accurately predict the frequency curve of a power system is solved, achieving fast and accurate frequency prediction and improving the frequency stability of the power system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEAST DIANLI UNIVERSITY
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional frequency response models struggle to predict power system frequency curves quickly and accurately, leading to a failure to implement timely frequency stability control and potentially causing system frequency collapse.
A frequency response model based on physical information neural networks is adopted. By designing the neural network structure, selecting an appropriate optimizer and loss function, and combining power system data for training and prediction, the power system frequency curve is generated.
It enables rapid and accurate prediction of power system frequency curves, improves frequency stability, and prevents system frequency collapse.
Smart Images
Figure CN122022010A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system frequency response models, and specifically to a frequency response model and method based on a physical information neural network. Background Technology
[0002] Currently, with the advancement of the dual-carbon strategy, vigorously developing new energy power generation has become an indispensable and important factor in accelerating carbon emission reduction, guiding green technology innovation, and continuously promoting the adjustment of industrial and energy structures. However, while improving the energy structure, it also brings problems such as weak system resilience, reduced inertia, and high uncertainty in new energy power generation. When the system is subjected to large disturbances, it may experience rapid frequency drops or even frequency collapse. Therefore, the ability to quickly and accurately predict power system frequency is of great significance for power system safety protection.
[0003] When power system operating conditions are complex and variable, traditional frequency response models struggle to predict power system frequency curves quickly and accurately, leading to delays in frequency stability control. Therefore, rapid and accurate prediction of power system frequency curves is crucial for improving power system frequency stability and preventing system frequency collapse. Summary of the Invention
[0004] To address the aforementioned issues, this invention proposes a frequency response model and method based on a physical information neural network, which can quickly and accurately predict the inertial center frequency curve of a power system under disturbances in different scenarios, thereby improving the frequency security and stability of complex power systems.
[0005] The frequency response model based on a physical information neural network proposed in this invention includes: reading and normalizing the power system frequency, unbalanced active power, and frequency response model parameters to ensure data consistency and comparability; designing the structure of the physical information neural network, including precisely configuring the number of neurons and connection methods in the input layer, intermediate layer, and output layer, and selecting a suitable optimizer, such as SGD, Adam, or RMSprop, to improve the efficiency and accuracy of network training; setting an appropriate learning rate to balance the convergence speed and generalization ability of the model during training; and determining the number of iterations to ensure that the network reaches the predetermined performance standard within a sufficient training period.
[0006] The loss function of the physical information neural network combines the physical term loss function with the data term loss function;
[0007] The physical term loss function is the mean square error between the frequency response model and the inertial center frequency of the power system, and the data term loss function is the mean square error between the neural network output value and the coefficients of the frequency response model.
[0008] Inverse normalize the feature values output by the neural network;
[0009] The features output from the aforementioned physical information neural network are input into the frequency response model. Through the model's calculations, a complete power system frequency curve is generated and output. The input data consists of the generator frequencies, system unbalanced active power, and frequency response model parameters for the first two seconds of the power system, which are then normalized for use in the training and prediction processes. The generator frequency sampling step size is 0.01. This processed data better suits the model's needs, improving its efficiency and performance.
[0010] The neural network employs a fully connected neural network architecture. The input layer is designed to receive 10 features, representing the frequency data of 10 generators over the first two seconds. The network contains 5 hidden layers, each with 64 neurons; the output layer contains 5 features; additionally, two transposed layers are embedded to ensure the correct shape of the data tensor during transmission between network layers and to ensure the accuracy of the final output format; the optimizer uses the Adam optimizer with a learning rate of 0.00004 and 20 iterations. This structural design aims to optimize the network's data processing capabilities and the accuracy of the output results.
[0011] The physical term loss function is the transfer function of the frequency response model: The output parameters A, B, C, D, and ΔP of the physical information neural network are substituted into the physical term loss function. In this way, the model generates a frequency response curve, which is compared with the inertial center frequency of the power system. The physical term loss value is obtained by calculating the mean squared error (MSE) between the two. The data term loss function evaluates the model's prediction accuracy by comparing the mean squared errors of the neural network's output parameters with those of the frequency response model parameters. Finally, the sum of the physical term loss function and the data term loss function constitutes the model's total loss function, which guides the model's training and optimization process.
[0012] The features output by the neural network are transformed back to their original scale from the normalized state. Specifically, the features are transformed back to their original scale from the normalized state.
[0013] The output parameters undergo denormalization, a process that reverses the normalization transformation previously applied to the training data, mapping the normalized features of the network output back to their actual values. This denormalization step ensures that the network's output can be directly compared with unprocessed real data and used for subsequent physical information analysis. The frequency response model based on the physical information neural network is then substituted into the transfer function of the frequency response model containing unknown parameters. Through this process, the model can generate frequency response curves, thereby enabling the prediction and analysis of the dynamic frequency of the power system.
[0014] To address the aforementioned technical problems, this invention also proposes a frequency response method based on a physical information neural network, comprising: Step 1, acquiring the frequency curves of each generator under different scenarios and disturbances, calculating the inertial center frequency of the power system, and using the active power disturbance of the power system as training data. A frequency response model is generated using the power system data, and the coefficients of the frequency response model are extracted as training data.
[0015] Step 2: Read the .csv file containing the response observations and normalize the training data. Normalization scales the training data to a uniform range. The processed data is used in the training and prediction processes. This process allows the data to better suit the model's needs, improving its efficiency and performance.
[0016] Step 3: Input the training data obtained in Step 2 into the physical information neural network for training.
[0017] Step 4: Input the test data into the trained model, and output the frequency response model parameters and the predicted value of unbalanced active power. Substitute the predicted parameters into the frequency response model to perform frequency prediction, and compare it with the actual inertial center frequency of the power system.
[0018] Beneficial effects
[0019] This invention uses the frequency, unbalanced power, inertial center frequency, and coefficients of the frequency response model of each generator in the power system for the first two seconds as training data; it constructs a physical information neural network; the frequency of each generator is used as input, and the coefficients and unbalanced power of the frequency response model are used as output. The physical information neural network is then optimized and trained through deep learning. By using the output parameters of the physical information neural network, the frequency response model can respond to the inertial center frequency of the power system under different operating conditions. This allows for the rapid understanding and analysis of dynamic frequency changes.
[0020] This invention addresses the difficulty in obtaining frequency response models for modern power systems under complex operating conditions, and is of great significance for improving power system frequency stability and preventing system frequency collapse. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings involved in the embodiments or the prior art are briefly described below. Obviously, these drawings illustrate several embodiments of the present invention, and those skilled in the art can derive other possible drawings based on these drawings without creative effort. The purpose of the drawings is limited to illustrating specific embodiments and does not limit the scope of the present invention.
[0022] Figure 1This is a flowchart of the frequency response model based on physical information neural networks proposed in this invention;
[0023] Figure 2 This is a 39-node system example used for demonstration purposes;
[0024] Figure 3 It is an image showing how the loss function values of the training and validation sets change with the number of iterations.
[0025] Figure 4 The coefficients of determination between the frequency curves predicted by the frequency response model based on the physical information neural network under different scenarios and the actual system frequency curves are presented in the form of a bar chart.
[0026] Figure 5 This paper presents a direct comparison of the predicted frequency curves of the inertial center frequency of the power system and the actual system frequency curves of the frequency response model based on physical information neural networks proposed in this invention under different scenarios. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.
[0028] The frequency response model based on physical information neural networks proposed in this invention includes:
[0029] Read the power system frequency information, unbalanced active power and frequency response model parameters as training data and normalize them to ensure data consistency and comparability;
[0030] Designing the structure of a physical information neural network involves precisely configuring the number of neurons and connections in the input, intermediate, and output layers, and selecting appropriate optimizers such as SGD, Adam, and RMSprop to improve the efficiency and accuracy of network training; setting a learning rate to balance the convergence speed and generalization ability of the model during training; and determining the number of iterations to ensure that the network reaches the predetermined performance standard within a sufficient number of training cycles.
[0031] The loss function of the physical information neural network combines the physical term loss function and the data term loss function; the physical term loss function is the mean square error between the frequency response model and the inertial center frequency of the actual power system, and the data term loss function is the mean square error between the neural network output value and the frequency response modulus coefficient.
[0032] The features output by the neural network are denormalized; the features output by the physical information neural network are input into the frequency response model, and the complete power system frequency curve is generated and output through the model's calculation.
[0033] The training data consists of the generator frequencies, unbalanced active power, and frequency response model parameters of the power system for the first two seconds, which are then normalized and used for training and prediction. The frequency sampling step size for the generators is 0.01.
[0034] The neural network adopts a fully connected neural network architecture; the input layer receives 10 features, contains 5 hidden layers with 64 neurons in each layer, the output layer outputs 5 features, and embeds two transposed layers and configures an Adam optimizer; the network optimizer uses the Adam optimizer with a learning rate of 0.00004 and 20 iterations; the structural design aims to optimize the network's data processing capability and the accuracy of the output results.
[0035] The frequency response method based on physical information neural networks proposed in this invention includes:
[0036] 1. Obtain the frequency curves of each generator under different scenarios and disturbances, calculate the inertial center frequency of the power system, and use the active power disturbance of the power system as training data. Generate a frequency response model using the power system data, and extract the coefficients of the frequency response model as training data.
[0037] 2. Read the .csv file containing response observations and normalize the power system frequency information and frequency response model data. Normalization scales the data to a uniform range. The processed data is used for training and prediction. This process allows the data to better suit the model's needs, improving its efficiency and performance.
[0038] 3. Construct a physical information neural network model.
[0039] The physical information neural network's input layer contains the frequency data of each generator in the power system for the previous two seconds. It has five hidden layers, each containing 64 neurons, and the output layer outputs the coefficients of the frequency response model and the unbalanced power. The neural network structure is a fully connected neural network. The loss function of the neural network is:
[0040]
[0041] Physical term loss function:
[0042] (A pred -A) 2 (2)
[0043] (B pred -B) 2 (3)
[0044] (C pred -C) 2 (4)
[0045] (D pred -D) 2 (5)
[0046] 4. Input the training data obtained in step two into the physical information neural network for training. The training process includes the following steps:
[0047] Step 1: The neural network receives training data and performs forward propagation to output the predicted values of A, B, C, D, and ΔP.
[0048] Step 2: Input the value output by the neural network into the loss function and compare the predicted value with the actual value.
[0049] Step 3: The physical information neural network performs backpropagation. The neural network updates the weights and biases based on the gradient of the loss function to optimize the model parameters.
[0050] Step 4: After each iteration, the neural network enters the validation phase, where only the loss function value of the current model is calculated without updating the weights, in order to evaluate the model's generalization ability.
[0051] Step 5: Repeat the above steps until the required number of iterations is completed to ensure that the model can learn and generalize effectively.
[0052] Step 6: After training is complete, plot the loss function values of the training and validation sets as a function of the number of iterations. By analyzing these plots, it is possible to determine whether the model is overfitting or underfitting, and adjust the model parameters or structure accordingly to optimize performance.
[0053] 5. Input the test data into the trained model, and the output is the frequency response model parameters and imbalance.
[0054] Active power prediction. The prediction parameters are substituted into the power system frequency response model to predict the frequency response, and then compared with the actual curve.
[0055] like Figure 1 As shown, the specific implementation process of a frequency response model method based on physical information neural networks is as follows:
[0056] S1. Obtain the frequency curves of each generator under different scenarios and disturbances, calculate the inertial center frequency of the power system, and use the active power disturbance of the power system as training data. Generate a frequency response model using the power system data, and extract the coefficients of the frequency response model as training data. Finally, normalize the training data.
[0057] S2. Train the physical information neural network. The neural network receives training data and outputs feature values. The loss function for the physical term is the transfer function of the frequency response model. Input of physical information neural network
[0058] The parameters A, B, C, D, and ΔP are substituted into the physical term loss function. In this way, the model generates a frequency response curve, which is compared with the actual inertial center frequency of the power system. The physical term loss value is obtained by calculating the mean squared error (MSE) between the two curves. The data term loss function evaluates the model's prediction accuracy by comparing the mean squared error between the neural network's output and the actual value. Finally, the sum of the physical term loss function and the data term loss function constitutes the model's total loss function. The neural network performs backpropagation based on the loss function, updating the neural network's weights and biases, thus enabling the model to undergo training and optimization.
[0059] S3. Input the test data into the trained neural network, and the physical information neural network outputs prediction parameters based on the input features.
[0060] S4: Substitute the predicted parameters output by the neural network into the frequency response model.
[0061] S5: Use a frequency response model to predict the frequency response after the disturbance and compare it with the actual system frequency curve.
[0062] Example Demonstration
[0063] To demonstrate the effect, a structure such as Figure 2 The 39-node power system shown has the following settings:
[0064] 1: Training data generation scenario: Synchronizer G03 exits operation, Synchronizer G05 exits operation, Synchronizer G07 exits operation, Synchronizer G08 exits operation, Synchronizer G10 exits operation.
[0065] 2: Data sampling step size: 0.01 seconds
[0066] 3: Number of training datasets: 3200 groups
[0067] 4: Fault occurrence time: 0 seconds;
[0068] 5: Observed electrical quantities: rotor angular velocity of each generator, system unbalanced power.
[0069] 6: Neural network parameters: 10 input features, 5 hidden layers, 64 hidden layer neurons, 5 output features, Adam optimizer, and learning rate of 0.0004.
[0070] Figure 3The graph shows the trend of the loss function values on the training and validation sets as the number of iterations changes.
[0071] As observed, with iteration, the loss function values of the training and validation sets gradually stabilize and become close to each other, indicating that the model's prediction error has reached a low level on both datasets without significant bias. This situation typically means that the physical information neural network is neither overfitting nor underfitting, demonstrating good generalization ability.
[0072] Figure 4 This section presents the coefficients of determination for predicting frequencies using a frequency response model based on physical information neural networks in different scenarios. The coefficient of determination is a statistic that measures the goodness of fit of a regression model. The value of the coefficient of determination is between 0 and 1; the closer the value is to 1, the better the model fits.
[0073] Figure 5 The invention presents a comparison between the predicted frequency curves of the power system inertial center frequency and the actual system frequency curves of the frequency response model based on physical information neural networks proposed in this invention under different scenarios, directly demonstrating the good fitting ability of the model.
[0074] Those skilled in the art should understand that the above embodiments are merely illustrative of the content of this disclosure and do not limit its scope. The system capacity, voltage, line parameters, etc., shown may vary depending on the specific circumstances of the power electronic grid-connected generator set and its grid connection. Based on this disclosure, those skilled in the art can make other changes or adjustments, and these changes still fall within the scope of this disclosure.
Claims
1. A frequency response model based on a physical information neural network, characterized in that, include: Read the power system frequency information, unbalanced active power and frequency response model parameters as training data and normalize them to ensure data consistency and comparability; The design of the physical information neural network structure includes precisely configuring the number of neurons and connection methods in the input, intermediate, and output layers, and selecting appropriate optimizers such as SGD, Adam, and RMSprop to improve the efficiency and accuracy of network training; setting a learning rate to balance the convergence speed and generalization ability of the model during training; determining the number of iterations to ensure that the network reaches the predetermined performance standard within a sufficient number of training cycles; and using a combination of physical term loss function and data term loss function for the loss function of the physical information neural network. The physical term loss function is the mean square error between the frequency response model and the inertial center frequency of the actual power system, while the data term loss function is the mean square error between the neural network output value and the frequency response modulus coefficient. The features output by the neural network are denormalized; the features output by the physical information neural network are input into the frequency response model, and the complete power system frequency curve is generated and output through the model's calculation.
2. The frequency response model based on a physical information neural network according to claim 1, characterized in that: The training data consists of the generator frequencies, unbalanced active power, and frequency response model parameters of the power system for the first two seconds, which are then normalized and used for training and prediction. The frequency sampling step size for the generators is 0.
01.
3. The frequency response model based on a physical information neural network according to claim 2, characterized in that: The neural network adopts a fully connected neural network architecture; the input layer receives 10 features, contains 5 hidden layers with 64 neurons in each layer, the output layer outputs 5 features, and embeds two transposed layers and configures an Adam optimizer; the network optimizer uses the Adam optimizer with a learning rate of 0.00004 and 20 iterations; the structural design aims to optimize the network's data processing capability and the accuracy of the output results.
4. The frequency response model based on a physical information neural network according to claim 3, characterized in that: The physical term loss function is the transfer function of the frequency response model: The output parameters A, B, C, D, and ΔP of the physical information neural network are substituted into the physical term loss function. The model generates a frequency response curve, which is compared with the inertial center frequency of the power system. The physical term loss value is obtained by calculating the mean square error between the two. The data term loss function evaluates the prediction accuracy of the model by comparing the mean square error between the output parameters of the neural network and the parameters of the frequency response model. The sum of the physical term loss function and the data term loss function constitutes the total loss function of the model.
5. The frequency response model based on a physical information neural network according to claim 4, characterized in that: The features output by the neural network are transformed from the normalized state back to their original scale. The output parameters are denormalized. By reversing the normalization transformation previously applied to the training data, the normalized features output by the network are mapped back to their actual values. The denormalization step ensures that the network output is directly compared with the unprocessed actual data and used for subsequent physical information analysis.
6. The frequency response model based on a physical information neural network according to claim 5, characterized in that: The physical information neural network outputs the frequency response model parameters and the predicted value of unbalanced active power, which are then substituted into the transfer function of the frequency response model. Through this process, the model can generate frequency response curves, thereby enabling the prediction and analysis of the dynamic frequency of the power system.
7. A method using a frequency response model based on a physical information neural network as described in any one of claims 1 to 6, characterized in that, include: Step 1: Obtain the frequency curves of each generator under different scenarios and disturbances, calculate the inertial center frequency of the power system, and use the active power disturbance of the power system as training data; use the power system data to generate a frequency response model, and extract the coefficients of the frequency response model as training data. Step 2: Read the .csv file containing the response observations and normalize the power system frequency information and frequency response model data; Normalization scales frequency response data to a uniform range; the processed data is used in the training and prediction processes; this processed data can better adapt to the needs of the model, improving the model's efficiency and performance. Step 3: Input the training data obtained in Step 2 into the physical information neural network for training; Step 4: Input the test data into the trained model, and output the frequency response model parameters and the predicted value of unbalanced active power; substitute the predicted parameters into the frequency response model to make frequency predictions and compare them with the actual inertial center frequency of the power system.